Hello, Luca is reporting that cgroups which have kvm instances inside never complete freezing. This can be trivially reproduced: root@test ~# mkdir /sys/fs/cgroup/test root@test ~# echo $fish_pid > /sys/fs/cgroup/test/cgroup.procs root@test ~# qemu-system-x86_64 --nographic -enable-kvm and in another terminal: root@test ~# echo 1 > /sys/fs/cgroup/test/cgroup.freeze root@test ~# cat /sys/fs/cgroup/test/cgroup.events populated 1 frozen 0 root@test ~# for i in (cat /sys/fs/cgroup/test/cgroup.threads); echo $i; cat /proc/$i/stack; end 2070 [<0>] do_freezer_trap+0x42/0x70 [<0>] get_signal+0x4da/0x870 [<0>] arch_do_signal_or_restart+0x1a/0x1c0 [<0>] syscall_exit_to_user_mode+0x73/0x120 [<0>] do_syscall_64+0x87/0x140 [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e 2159 [<0>] do_freezer_trap+0x42/0x70 [<0>] get_signal+0x4da/0x870 [<0>] arch_do_signal_or_restart+0x1a/0x1c0 [<0>] syscall_exit_to_user_mode+0x73/0x120 [<0>] do_syscall_64+0x87/0x140 [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e 2160 [<0>] do_freezer_trap+0x42/0x70 [<0>] get_signal+0x4da/0x870 [<0>] arch_do_signal_or_restart+0x1a/0x1c0 [<0>] syscall_exit_to_user_mode+0x73/0x120 [<0>] do_syscall_64+0x87/0x140 [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e 2161 [<0>] kvm_nx_huge_page_recovery_worker+0xea/0x680 [<0>] kvm_vm_worker_thread+0x8f/0x2b0 [<0>] kthread+0xe8/0x110 [<0>] ret_from_fork+0x33/0x40 [<0>] ret_from_fork_asm+0x1a/0x30 2164 [<0>] do_freezer_trap+0x42/0x70 [<0>] get_signal+0x4da/0x870 [<0>] arch_do_signal_or_restart+0x1a/0x1c0 [<0>] syscall_exit_to_user_mode+0x73/0x120 [<0>] do_syscall_64+0x87/0x140 [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e The cgroup freezing happens in the signal delivery path but kvm_vm_worker_thread() thread never call into the signal delivery path while joining non-root cgroups, so they never get frozen. Because the cgroup freezer determines whether a given cgroup is frozen by comparing the number of frozen threads to the total number of threads in the cgroup, the cgroup never becomes frozen and users waiting for the state transition may hang indefinitely. There are two paths that we can take: 1. Make kvm_vm_worker_thread() call into signal delivery path. io_wq_worker() is in a similar boat and handles signal delivery and can be frozen and trapped like regular threads. 2. Keep the count of threads which can't be frozen per cgroup so that cgroup freezer can ignore these threads. #1 is better in that the cgroup will actually be frozen when reported frozen. However, the rather ambiguous criterion we've been using for cgroup freezer is whether the cgroup can be safely snapshotted whil frozen and as long as the workers not being frozen doesn't break that, we can go for #2 too. What do you guys think? Thanks. -- tejun
